Statistical Toolkit

Statistical methods form the necessary scientific basis for assessing trends in climate, seismicity, and the likelihood of extreme events and associated hazards.

Introduction
Traditional methods used for most KNMI products and information services assume a stationary geophysical environment. These methods seem no longer prudent in times of climate change and increased induced seismicity. The suite of statistical methods that is used at KNMI for our key products (including the climate atlas, climate scenarios and seismic hazard assessments) is due to be renewed.

Current sub-projects

Refined models of the statistics of extreme weather events
New models of the statistics of extreme events, specifically designed for extrapolation to very rare  events, are evaluated for application to extreme wind and precipitation. Software for estimation of these models from data is developed in R and distributed as EVTools

Long-range dependence in weather extremes
In current models of the statistics of extreme weather, dependence among severe weather events occurring in different years is considered absent, and time-dependence is limited to a deterministic trend in the climatology. However, close examination of wind measurements reveals dependence on all time scales longer than a year. By modelling this long-range dependence, we can improve estimation, trend analysis and uncertainty assessment.

Advanced calibration of weather data
Quantile mapping is an effective tool for correcting bias in measurements, e.g. after relocation of a weather station, or bias in model predictions. However, it corrects one variable at a time, independent from all the others, which may lead to inconsistencies. A novel approach based on optimal transport theory offers the potential to overcome this limitation.

Completed  sub-projects

Fog detection from camera images (with KNMI DataLab)
The idea of this project is to unearth the potential of image data, e.g. provided by traffic cams, to expand the observation network. With the expanded observation basis, nowcasting and the validation of the KNMI weather models can be improved. Algorithms are found in visDec.

Extreme temperatures for KNMI forecasts
The KNMI forecasts plume is being enriched by climate information. In this project we produced return levels of daily temperatures which take the annual cycle into account. The corresponding R-package is knmipluim.

Monotone trend in the GPD scale parameter (with TU Delft)
We developed an algorithm to efficiently estimate a monotone trend in the scale parameter of the Generalized Pareto distribution, which is then applied to the very long Central England temperature series. The corresponding R-package is gpdIcm.

Peak values of the Central England temperature series.
Peak values of the Central England temperature series. The blue lines correspond to linear 0.5, 0.75, and 0.975 regression quantiles. The red lines are these quantiles from the monotone scale GPD approach. The dashed red line shows the 100-y return level.
Trend in mean annual temperature for selected cities in Europe, based on the EOBS data set. An interesting feature is that only the Dutch cities show a distinction between mean and median.
Trend in mean annual temperature for selected cities in Europe, based on the EOBS data set. An interesting feature is that only the Dutch cities show a distinction between mean (red) and median (blue).

Generalized additive modelling for mean temperatures
For the modelling of trends in extreme mean temperatures, e,g. monthly means, we are exploring  the use of generalized additive modelling.

Poisson tracking of earthquake activity (with University of Amsterdam)
We are interested in Poisson tracking of the earthquake activity in Groningen, which is more flexible than classical regression approaches.